Douglas Turnbull, Luke Barrington, Gert Lanckriet Computer

Five Approaches to Collecting Tags for Music
Approach
Survey
Social Tags
Summary
Experts are paid to annotate
songs using a stardard form
Large community contributes tags
using a social network
Players produce tags as they play
a video game
Strengths
Custom-tailored vocabulary
High-quality annotations
Strong labeling
Collective wisdom of crowds
Unlimited vocabulary
Provides social context
Collective wisdom of the crowds
Fast paced for rapid data collection
Large corpus of publicly-available documents
Small, predetermined vocabulary
Human-labor intensive
Ad-hoc
aasdf annotation behavior
Produces weak labeling
Sparse/missing data in long tail
Weaknesses
Time consuming approach lacks scalability
Game Tags
Webtags
Paid 55 undergraduates to annotate 500
songs by 500 artists using a vocabulary of
tags. Each song was annoated my a minimum of 3 individuals.
This data serves as the ground truth.
There are 87 “long tail” songs from Magnatunes.
All Songs
Long Tail
The vocabulary consists of 109 tags that
relate to genre, instrumentation, emotion,
usage, rhythm, vocals, and other musical
characteristics.
Density
AUC-ROC
Top 10 Prec
AUC-ROC
Top 10 Prec
Hybrid
Combination of approaches
Entertaining incentives produce high-quality tags
No direct human involvement
Provides social context
Not affected by cold-start problem
No direct human involvement
Produces strong labeling
Combine social context and audio content
Use strengths, remove weaknesses
Multi-tiered approach to cold-start problem
“Gaming”
the system
aasdf
Difficult to create viral gaming experience
Based on short clips, rather than songs
Text-mining
introduces noise
aasdf
Produces weak labeling
Sparce/missing data in long tail
Computationally
intensive
aasdf
Limited by training data
Based solely on audio content, no context
Increased
system complexity
aasdf
Combining data sources can be tricky
Semantic
Music
Discovery
Engine
CAL 500 Data Set [1]
Density
Autotags
Annotate audio content using signal
processing and machine learning
Analyze a corpus of music reviews,
artist bios, blogs, discussion boards
Example
Algorithm
Douglas Turnbull, Luke Barrington, Gert Lanckriet
Computer Audition Laboratory, UC San Diego
Audioscrobbler
Attempted to collect a list of tags associated with each CAL500 song and each
CAL500 artist from Last.fm’s Audioscrobbler website. For each song, the song
and artist lists were combined.
ListenGame [2]
During a two week pilot study of ListenGame, we collected 16,500 tags for 250
of the CAL500 songs from 440 players.
Each of the 27,250 song-tag pairs were
present 1.8 times on average.
The combined list was matched to the
CAL500 vocabulary. We attempted to use
synonyms, alternative spellings, and wildcard matching to improve coverage.
Relevance Scoring [3]
1) Collect Corpus
query google with song, artist and album
2) Query Corpus with Tag
find most relevant song given tag
Site-Specific
use web documents from sites that are
known to have high-quality content (Rolling
Stones, AMG AllMusic, etc
Weighted RS
weight pages by tag relevance
1.00
1.00
0.97
0.23
0.62
0.37
1.00
1.00
0.57
0.03
0.54
0.19
Supervised Multilabel Model [1]
MFCC+Delta Feature Space
One GMM per tag
Mixture Hiearchies EM to train GMMs
Produces “Semantic Multinomial”
distribution over tag vocabulary for
each novel song
Rank-Based Interleaving
Given a tag, rank songs based on
their best rank according to other appoaches
Other hybrid approachs
1) Kernel Combination [4]
2) RankBoost [5]
3) Calibarated Score Averaging [5]
Top performing system in 2008
MIREX Audio Tag Classification Task
0.37
0.65
0.32
0.67
0.66
0.32
1.00
0.69
0.33
1.00
0.74
0.38
Based on our experimental setup, the
long tail results are misleading because
the selection of songs in ListenGame is
independent of popularity.
0.25
0.56
0.18
1.00
0.70
0.27
1.00
0.71
0.28
Please Contact: Douglas Turnbull <[email protected]>
Data, Papers, and additional Information can be found at: http://cosmal.ucsd.edu/cal/
[1] Turnbull, Barrington, Torres, Lanckriet Semantic Annotation and Retrieval Music and Sound Effects. TASLP 2008
[2] Turnbull, Liu Barrington, Lanckriet Using Games to Collect Semantic Information About Music. ISMIR 2007
[3] Knees, Pohle, Schedl, Schnitzer, Seyerlehner A Document-Centered Approach to a Natural Language Music Search Engine. ECIR 2008
[4] Barrington, Yazdani, Turnbull, Lanckriet Combining Feature Kernels for Semantic Music Retrieval. ISMIR 2008
[5] Turnbull Design and Development of a Semantic Music Discovery Engine. Ph.D. Thesis, UC San Diego 2008